Add kernel profiles (148k rows), workload traces, serving predictions, roofline quadrant data

README.md

@@ -12,7 +12,7 @@ tags:
   - sglang
   - agentic-workloads
 size_categories:
-  - 1K<n<10K
+  - 100K<n<1M
 pretty_name: AgentPerfBench
 version: "1.0"
 configs:
@@ -24,6 +24,22 @@ configs:
     data_files:
       - split: summary
         path: distributional/summary.parquet
+  - config_name: kernel_profiles
+    data_files:
+      - split: kernels_labeled
+        path: kernel_profiles/kernels_labeled.parquet
+      - split: roofline_quadrant
+        path: kernel_profiles/roofline_quadrant.parquet
+  - config_name: workload_traces
+    data_files:
+      - split: coding_agent_prompts
+        path: workload_traces/coding_agent_prompts.parquet
+      - split: osworld_trajectories
+        path: workload_traces/osworld_trajectories.parquet
+  - config_name: predictions
+    data_files:
+      - split: serving_predictions
+        path: predictions/serving_predictions.parquet
 dataset_info:
   - config_name: trace_replay
     features:
@@ -169,11 +185,11 @@ dataset_info:
 
 # AgentPerfBench
 
-LLM inference benchmark: 3,392 runs measuring TTFT, TPOT, ITL, and throughput across 9 models, up to 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
+LLM inference benchmark: 3,392 serving runs, 148,077 per-kernel CUDA profiles, 4,715 latency predictions, and 560 workload traces across 9 models, up to 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
 
 ## Dataset configurations
 
-The dataset provides two configurations. *trace_replay* replays exact input/output sequences from recorded agent sessions. *distributional* samples from statistical distributions fitted to those same workloads, trading fidelity for faster sweeps across the hardware matrix.
+Five configurations covering serving benchmarks, kernel profiling, workload traces, and latency predictions.
 
 ### trace_replay (3,147 rows)
 
@@ -187,6 +203,26 @@ Samples ISL/OSL from lognormal distributions fitted to real workload statistics.
 
 6 profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
 
+### kernel_profiles (148,077 + 2,163 rows)
+
+Per-kernel CUDA profiling data from NCU (Nsight Compute). Two splits:
+
+- **kernels_labeled** (148,077 rows): Individual kernel invocations across 4 GPUs (A100, H100, RTX 3090, RTX 2080Ti) and 13 model/sweep sources. Columns include kernel_family, kernel_name, M/N/K dimensions, gpu_time_duration_ms, dram_bytes_sum, launch_block_size, launch_grid_size, and register pressure.
+- **roofline_quadrant** (2,163 rows): Operational intensity and achieved throughput per kernel, for roofline analysis. H100 reference hardware (989 peak TFLOPS, 3.35 TB/s HBM).
+
+### workload_traces (500 + 60 rows)
+
+Raw agent session recordings used to derive the trace_replay profiles.
+
+- **coding_agent_prompts** (500 rows): System/user prompt pairs with output token counts from SWE-Bench coding agent sessions.
+- **osworld_trajectories** (60 rows): Multi-turn OSWorld sessions with per-turn action/observation data (up to 30 turns per session).
+
+Note: SWE-Bench (1.6 GB) and TerminalBench (1.9 GB) trajectory files are available on the project's R2 storage but excluded from this release due to size.
+
+### predictions (4,715 rows)
+
+Predicted vs. measured latency for each serving configuration. Columns include ttft_pred/ttft_meas/ttft_err, tpot_pred/tpot_meas/tpot_err, e2el_pred/e2el_meas/e2el_err, plus cache-aware prediction metadata (cache_hit_rate, cache_aware_applied, multiturn_prediction_mode). Covers 14 hardware configs across all models and profiles.
+
 ### Concurrency filtering
 
 The benchmark harness capped actual concurrent connections at the session pool size. Rows where declared concurrency exceeded the pool were excluded:
@@ -274,8 +310,9 @@ Each row in `summary.parquet` (both configs):
 ```python
 from datasets import load_dataset
 
+# Serving benchmark results
 ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
-# or "distributional"
+# or "distributional", "kernel_profiles", "workload_traces", "predictions"
 ```
 
 ## Benchmark methodology
@@ -291,7 +328,7 @@ ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
 ## Future releases
 
 - Per-request and multi-turn granularity data (pending raw JSON availability from collection infrastructure).
-- Per-kernel CUDA roofline profiles (PyTorch profiler, 2-layer forward passes, batch sizes 1/4/8/32/64).
+- SWE-Bench (1.6 GB) and TerminalBench (1.9 GB) trajectory files.
 - This is version 1.0. Updates will be tagged with semantic versions.
 
 ## Intended uses

kernel_profiles/kernels_labeled.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3158723dcd2840b2c0080dbfc4d91e8004d7629743b679c9f970564099eb3649
+size 656666

kernel_profiles/roofline_quadrant.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3b06c80f99a3d771dac29e0e329678d2c966ad691959a12b504098e8076478b
+size 64944

predictions/serving_predictions.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11041092d91db7a96816bb61ebb092987dab5bfa2f7ae437b220d5131245b051
+size 3186089

workload_traces/coding_agent_prompts.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0de8f6b528531b69edb19099a6afd3d66f47195fa5dc8108f91943a187beedd
+size 390896

workload_traces/osworld_trajectories.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbe88d7d9e4bea71468c6732ff93c519a201deb2f456b6452eaf5fae5510aa16
+size 3443036